Velocirag

What is Velocirag?

Velocirag is a lightweight, local RAG (Retrieval-Augmented Generation) system for AI agents. It combines four retrieval methods — vector similarity, BM25 keyword matching, knowledge graph traversal, and metadata filtering — fused through reciprocal rank fusion with cross-encoder reranking, all running on ONNX Runtime without PyTorch or a GPU. It includes an MCP server for agent integration, a Unix socket daemon for warm queries, and a CLI.

How to use Velocirag?

Install with pip install "velocirag[mcp]", index documents with velocirag index ./my-docs, then start the MCP server with velocirag mcp. Configure the MCP server in Claude, Cursor, or Windsurf using the provided JSON snippets. Alternatively, use the Python API (Embedder, VectorStore, Searcher) or the search daemon (velocirag serve) for warm queries.

Key features of Velocirag

• ONNX Runtime, no PyTorch, no GPU required
• Four-layer fusion: vector, keyword, knowledge graph, metadata
• Cross-encoder reranking via TinyBERT (included)
• Incremental graph updates with file-centric provenance
• MCP server with five tools: search, index, add_document, health, list_sources
• Search daemon keeps model warm over Unix socket
• CPU-only, <8GB RAM, no API keys or external services

Use cases of Velocirag

• AI agents needing fast, local RAG without external dependencies
• Semantic search over technical documentation with metadata filters
• Knowledge management for evolving document sets with incremental updates
• Integration into Claude, Cursor, or Windsurf via the MCP server

FAQ from Velocirag

How does Velocirag differ from other RAG tools like LangChain or Chroma?

Velocirag offers four retrieval layers (vector, keyword, graph, metadata) compared to one or two in most alternatives. It includes cross-encoder reranking, incremental updates, and a built-in MCP server, all without requiring PyTorch, a GPU, or an LLM for search. Install size is ~80MB versus 750MB+ for LangChain or LlamaIndex.

Does Velocirag require an LLM or API keys?

No. Velocirag runs fully locally with no LLM required for search and no API keys. The embedding and reranking models (MiniLM-L6-v2, TinyBERT) are downloaded on first use via ONNX